Xeon Phi's had that - up to 16GB of ridiculously fast memory in-package and the ability to drive external memory at more pedestrian speeds.

IIRC the 16GB could be configured as an "L4 cache" for the rest of the memory.

I assume the same approach could be taken.

OTOH, it shouldn't be too complicated to just add more memory dies to the package as higher densities become available.